🚨All attention refactor🚨 (#35235)

* refactor LlamaAttention

* minimal changes

* fix llama

* update

* modular gemmas

* modular nits

* modular updates

* nits

* simplify

* gpt2

* more modualr and fixes

* granite

* modular modular modular

* nits

* update

* qwen2 + starcoder2

* mostly gemma2

* Update image_processing_auto.py

* fix

* Update modular_starcoder2.py

* fix

* remove all copied from attentions

* remove gcv

* make fix-copies

* oups

* oups2.0

* fix some modulars + all copied from

* should be good now

* revert unwanted changes

* Update modeling_decision_transformer.py

* finish cleanup

* Update modeling_olmo.py

* consistency

* re-add gradient checkpointing attribute

* fix

* style

* make config necessary

* bis

* bis

* Update modeling_my_new_model2.py

* is_causal attr

* fix

* remove past kv return from decoder layer

* fix

* default rope config

* correctly fix rope config

* fix bias

* fix gpt2 attention output

* fix test

* fix inits

* fix default sdpa

* fix default sdpa implementation

* harmonize classes

* fix mistral

* fix sliding window models

* mixtral

* be more explicit

* style

* fix

* several fixes

* Update modeling_dbrx.py

* fix test

* olmo + phi

* rotary

* syle

* phi

* phi again

* again

* kwargs

* Update test_modeling_common.py

* skip fx tracing tests

* Update modeling_utils.py

* gemma 2

* again

* Update modeling_recurrent_gemma.py

* gemma2

* granite

* style

* starcoder

* Update sdpa_attention.py

* switch args

* Update modeling_mllama.py

* fix

* cache type tests

* gpt2

* Update test_modeling_common.py

* fix

* consistency

* fix shape with encoder

* should be the last one

* tests non model

* most comments

* small oupsi

* be more explicit in modulars

* more explicit modulars

* CIs! it works locally

* add kwargs to _flash_attention_forward

---------

Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>

tests/test_modeling_flax_common.py

@@ -23,6 +23,7 @@ import numpy as np
 
 import transformers
 from transformers import is_flax_available, is_torch_available
+from transformers.cache_utils import DynamicCache
 from transformers.models.auto import get_values
 from transformers.testing_utils import CaptureLogger, is_pt_flax_cross_test, require_flax, torch_device
 from transformers.utils import CONFIG_NAME, GENERATION_CONFIG_NAME, logging
@@ -180,7 +181,7 @@ class FlaxModelTesterMixin:
             check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
 
     # (Copied from tests.test_modeling_common.ModelTesterMixin.check_pt_flax_outputs)
-    def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
+    def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-4, name="outputs", attributes=None):
         """
         Args:
             model_class: The class of the model that is currently testing. For example, ..., etc.
@@ -190,7 +191,6 @@ class FlaxModelTesterMixin:
                 Currently unused, but in the future, we could use this information to make the error message clearer
                 by giving the name(s) of the output tensor(s) with large difference(s) between PT and Flax.
         """
-
         self.assertEqual(type(name), str)
         if attributes is not None:
             self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")
@@ -235,6 +235,8 @@ class FlaxModelTesterMixin:
                 attributes = tuple([f"{name}_{idx}" for idx in range(len(fx_outputs))])
 
             for fx_output, pt_output, attr in zip(fx_outputs, pt_outputs, attributes):
+                if isinstance(pt_output, DynamicCache):
+                    pt_output = pt_output.to_legacy_cache()
                 self.check_pt_flax_outputs(fx_output, pt_output, model_class, tol=tol, name=attr)
 
         elif isinstance(fx_outputs, jnp.ndarray):